Fix FT malloc size for beam search. (#3662)

sserdoubleh · FrostML · web-flow · commit 58a5a9dcfd94 · 2022-11-04T19:28:09.000+08:00
Co-authored-by: liu zhengxi &lt;380185688@qq.com&gt;
diff --git a/paddlenlp/ops/patches/FasterTransformer/fastertransformer/decoding_beamsearch.h b/paddlenlp/ops/patches/FasterTransformer/fastertransformer/decoding_beamsearch.h
@@ -275,8 +275,8 @@ class DecodingBeamsearch {
     }
 
     size_t lm_head_buffer_size = (prefix_lm)
-                                     ? decoder_normed_result_buffer_size
-                                     : decoder_normed_result_buffer_size * 3;
+                                     ? decoder_normed_result_buffer_size * 3
+                                     : decoder_normed_result_buffer_size;
 
     size_t datatype_buf_size =
         from_tensor_size * 2 + decoder_workspace_size +

Original file line number	Diff line number	Diff line change
`@@ -275,8 +275,8 @@ class DecodingBeamsearch {`
`275`	`275`	`}`
`276`	`276`
`277`	`277`	`size_t lm_head_buffer_size = (prefix_lm)`
`278`		`- ? decoder_normed_result_buffer_size`
`279`		`- : decoder_normed_result_buffer_size * 3;`
	`278`	`+ ? decoder_normed_result_buffer_size * 3`
	`279`	`+ : decoder_normed_result_buffer_size;`
`280`	`280`
`281`	`281`	`size_t datatype_buf_size =`
`282`	`282`	`from_tensor_size * 2 + decoder_workspace_size +`